TypeSafe AI founder Diogo Almeida on a viral launch he calls 'above the 100th percentile' — and why structured AI decisions are the missing layer
Key Points
- TypeSafe AI's Jev product, which builds structured decision models for automation rather than generative text, achieved a launch response Almeida describes as 'above the hundredth percentile,' with demand outpacing anticipated infrastructure needs.
- Jev prioritizes calibration and uncertainty signaling over confident predictions, allowing systems to safely automate workflows like data entry and customer service classification that typically require human judgment.
- TypeSafe AI has raised less than $40 million and claims profitability through a fundamentally different compute architecture, avoiding the frontier compute spending typical of AI labs.
Summary
TypeSafe AI is building what its founder calls the missing layer between AI models and production automation: structured decision and classification models, branded Jev, designed for reliability and calibration rather than generative text output. Diogo Almeida, the company's CEO, came out of OpenAI's post-training work and says the core idea seemed so obvious to anyone at the intersection of developer and researcher that he assumed Anthropic was already building it. They weren't, or at least hadn't shipped it.
The launch has been, by Almeida's own admission, well above anything he planned for. He describes it as "above the hundredth percentile," with comparisons to ChatGPT's developer impact circulating despite Jev being a fundamentally different category of product. Demand has outpaced infrastructure on dimensions he says he had never anticipated.
“People are calling it like a ChatGPT for developers. And that shouldn't even be possible because ChatGPT is a consumer product. / Make AI actually useful for automation. AI over-promise, under-deliver has been a major issue for years, and it's been automating almost nothing. / You can't get to Neolab compute with less than $40,000,000. Yet we have [built a profitable AI company] because we are just innovating.”
What Jev does
The pitch is that AI has overpromised and underdelivered on automation for years, and the core reason is that general-purpose LLMs don't play nicely with production code. Jev doesn't generate strings. It won't write code or address coding agents, which Almeida frames as a deliberate choice rather than a gap. The target is the "boring automation" that still requires a human: data entry, accounting workflows, customer service classification.
The calibration point is central to his argument. An LLM might perform at a high level 95% of the time and fail badly the other 5%, but if it can't tell you which mode it's in, you can't safely automate anything that normally requires human judgment. Jev is designed to signal uncertainty rather than confidently hallucinate, which Almeida argues is table stakes for unsupervised automation.
The classifier debate
The "it's just a classifier" criticism from the ML community doesn't bother Almeida. Classifiers, he argues, dominated pre-generative AI precisely because they were built for usefulness in systems, not for demos. Jev draws on that tradition while discarding the ML 1.0 concepts that don't transfer to today's models. The launch, he says, was aimed at developers, not researchers, which is why researcher pushback largely misses the point.
Compute and capital
TypeSafe AI has raised less than $40 million, well short of what a typical AI lab would need to compete on frontier compute. Almeida says the company is profitable, which he attributes to a fundamentally different compute architecture rather than cost discipline on a standard stack. The bottleneck changes daily as the system scales into demand it wasn't designed for.
Every deal, every interview. 5 minutes.
TBPN Digest delivers summaries of the latest fundraises, interviews and tech news from TBPN, every weekday.